Modeling software reveals patterns in continuous seismic waveforms during a series of magnitude 5 stick-slip earthquakes

A team at Los Alamos National Laboratory has used machine learning — an application of artificial intelligence — to detect the hidden signals that precede an earthquake. The findings at Hawaii’s Kīlauea volcano are part of years of research pioneered at Los Alamos, and this latest study is the first time scientists have been able to detect these warning signals on a stick-slip fault that can cause massive destruction.

The article appeared in the magazine Geophysical Research Letters.

“We wanted to see if we could filter out signals from the noise and determine what stage of the stress cycle the system was in in terms of approaching a major earthquake-causing landslide,” said Christopher Johnson, a seismologist at Los Alamos and lead researcher on the team. “This is the first time we’ve been able to use this method on an earthquake of this type and magnitude.”

The team used data recorded by the U.S. Geological Survey’s Hawaiian Volcano Observatory between June 1 and August 2, 2018. During that time, the volcano experienced more than 50 quakes of varying magnitudes. The researchers focused on 30-second windows of seismic data and their model identified something like a fingerprint, a hidden signal that tracked the stress cycle of each event. On average, this hidden signal appeared continuously before any detectable large ground motion occurred.

Combined with previous tests, the results suggest that some earthquake faults have a similar physical nature, meaning this method could be used to assess earthquake hazards worldwide.

Patterns in the noise

The research builds on previous work from Los Alamos on faults in California and the Pacific Northwest, where machine learning was able to detect these precursor signals.

As tectonic plates push against each other, they create weak tremors in the ground called continuous acoustic or seismic emissions. These signals appear like waveforms when recorded, but were previously thought to be noise – data with no information about the state of the fault. Instead, Los Alamos researchers have found that continuous acoustic emission waveforms are actually rich in data and can be used to infer physical properties of a fault, such as displacement, friction and thickness.

Most importantly, Los Alamos scientists discovered highly predictable patterns in the signals, a kind of timeline to failure.

“By looking at these continuous signals, we can filter out information that tells us where the fault is in its charge cycle,” Johnson said. “We watch the noise evolve, and that gives us details about its current state and where it is in the slip cycle.”

From slow slip to stick slip

In their research, the team successfully applied the approach for the first time to seismogenic faults, the layer where earthquakes originate. In this case, it was a sequence of highly active magnitude 5 stick-slip events at Kīlauea volcano, which experienced a months-long seismic event that caused the caldera to subside by 1,600 feet.

During this time, a global navigation satellite system measured ground displacements with millimeter precision. The machine learning model analyzed this data, processed the seismic signals, and successfully estimated the ground displacement and the time to the next fault.

Previously, Los Alamos researchers had applied similar machine learning models to slow-slip events, which cause the ground to shake gently for days, months or even years before a seismic event. Such large data sets were helpful in training the machine learning models. But the most destructive earthquakes are caused by stick-slip faults, like those found at Kīlauea volcano, which can generate much stronger ground motions much more quickly and have so far eluded prediction.